Depends. The only predictions I have seen here are the centaurs vs anti centaurs of Doctorow, and even his analysis I find pretty flimsy.

I dont think the race to shove an LLM into everything is going to grow the pie.

But I also dont think it is impossible that a use case will present itself that will create further jobs.

The issue is that its largely unpredictable.

Its a bit like, we are sitting around in the 1950s trying to predict how computers will affect the economy.

It is going to take more than 1 successful deductive leap to get us from 1950s computing -> miniaturisation -> computer in every home -> internet communications.

Every deductive leap we take is extremely prone to being wrong.

We simply cannot lie back and imagine every productive relationship in the economy and then extrapolate every centaur and anti centaur possible for it.

What we do know is that theres a bit of a gold rush to effectively brute force every possible AI variant into every productive relationship in the economy. The fastest way to get the answer to your question is to do it. Possibly the only way to get the answer is to do it.

For instance, someone might imagine LLMs simply eating a whole bunch of service industry jobs. At the same time, theres a mid state where it eats some, but the remaining staff are employed to monitor the LLMs to prevent them handing out free shit to smart shoppers. Its also easy enough to imagine that LLMs never quite get there and the risk is too large for foul play, so they just dont gain that kind of traction. Its also possible to imagine an end state where LLMs can get to 0% risk if they are constantly trained on human data coming from humans doing the same job, and that humans are gainfully employed in parallel with LLMs. Its possible that LLMs are great at business as usual, but the risk emerges when company policies change, and the cost of retraining LLMs makes it impractical for move fast and break things companies to do anything but hire humans. My favourite scenario is one where humans are largely AI assisted, trained on particular people, and theres a massive cybercrime industry built around exfiltrating LLM training weights trained on high functioning humans and deploying them without humans to the third world to help them get 80% of the quality of first world businesses, making them heavily competitive.

We dont know what we dont know.